Distributed RPC Framework¶

The distributed RPC framework provides mechanisms for multi-machine model training through a set of primitives to allow for remote communication, and a higher-level API to automatically differentiate models split across several machines.

Warning

APIs in the RPC package are stable. There are multiple ongoing work items to improve performance and error handling, which will ship in future releases.

Warning

CUDA support was introduced in PyTorch 1.9 and is still a beta feature. Not all features of the RPC package are yet compatible with CUDA support and thus their use is discouraged. These unsupported features include: RRefs, JIT compatibility, dist autograd and dist optimizier, and profiling. These shortcomings will be addressed in future releases.

Note

Please refer to PyTorch Distributed Overview for a brief introduction to all features related to distributed training.

Basics¶

The distributed RPC framework makes it easy to run functions remotely, supports referencing remote objects without copying the real data around, and provides autograd and optimizer APIs to transparently run backward and update parameters across RPC boundaries. These features can be categorized into four sets of APIs.

Remote Procedure Call (RPC) supports running a function on the specified destination worker with the given arguments and getting the return value back or creating a reference to the return value. There are three main RPC APIs: rpc_sync() (synchronous), rpc_async() (asynchronous), and remote() (asynchronous and returns a reference to the remote return value). Use the synchronous API if the user code cannot proceed without the return value. Otherwise, use the asynchronous API to get a future, and wait on the future when the return value is needed on the caller. The remote() API is useful when the requirement is to create something remotely but never need to fetch it to the caller. Imagine the case that a driver process is setting up a parameter server and a trainer. The driver can create an embedding table on the parameter server and then share the reference to the embedding table with the trainer, but itself will never use the embedding table locally. In this case, rpc_sync() and rpc_async() are no longer appropriate, as they always imply that the return value will be returned to the caller immediately or in the future.
Remote Reference (RRef) serves as a distributed shared pointer to a local or remote object. It can be shared with other workers and reference counting will be handled transparently. Each RRef only has one owner and the object only lives on that owner. Non-owner workers holding RRefs can get copies of the object from the owner by explicitly requesting it. This is useful when a worker needs to access some data object, but itself is neither the creator (the caller of remote()) or the owner of the object. The distributed optimizer, as we will discuss below, is one example of such use cases.
Distributed Autograd stitches together local autograd engines on all the workers involved in the forward pass, and automatically reach out to them during the backward pass to compute gradients. This is especially helpful if the forward pass needs to span multiple machines when conducting, e.g., distributed model parallel training, parameter-server training, etc. With this feature, user code no longer needs to worry about how to send gradients across RPC boundaries and in which order should the local autograd engines be launched, which can become quite complicated where there are nested and inter-dependent RPC calls in the forward pass.
Distributed Optimizer’s constructor takes a Optimizer() (e.g., SGD(), Adagrad(), etc.) and a list of parameter RRefs, creates an Optimizer() instance on each distinct RRef owner, and updates parameters accordingly when running step(). When you have distributed forward and backward passes, parameters and gradients will be scattered across multiple workers, and hence it requires an optimizer on each of the involved workers. Distributed Optimizer wraps all those local optimizers into one, and provides a concise constructor and step() API.

RPC¶

Before using RPC and distributed autograd primitives, initialization must take place. To initialize the RPC framework we need to use init_rpc() which would initialize the RPC framework, RRef framework and distributed autograd.

torch.distributed.rpc.init_rpc(name, backend=None, rank=- 1, world_size=None, rpc_backend_options=None)[source]¶

Initializes RPC primitives such as the local RPC agent and distributed autograd, which immediately makes the current process ready to send and receive RPCs.

Parameters

name (str) – a globally unique name of this node. (e.g., Trainer3, ParameterServer2, Master, Worker1) Name can only contain number, alphabet, underscore, colon, and/or dash, and must be shorter than 128 characters.
backend (BackendType, optional) – The type of RPC backend implementation. Supported values is BackendType.TENSORPIPE (the default). See Backends for more information.
rank (int) – a globally unique id/rank of this node.
world_size (int) – The number of workers in the group.
rpc_backend_options (RpcBackendOptions, optional) – The options passed to the RpcAgent constructor. It must be an agent-specific subclass of RpcBackendOptions and contains agent-specific initialization configurations. By default, for all agents, it sets the default timeout to 60 seconds and performs the rendezvous with an underlying process group initialized using init_method = "env://", meaning that environment variables MASTER_ADDR and MASTER_PORT need to be set properly. See Backends for more information and find which options are available.

The following APIs allow users to remotely execute functions as well as create references (RRefs) to remote data objects. In these APIs, when passing a Tensor as an argument or a return value, the destination worker will try to create a Tensor with the same meta (i.e., shape, stride, etc.). We intentionally disallow transmitting CUDA tensors because it might crash if the device lists on source and destination workers do not match. In such cases, applications can always explicitly move the input tensors to CPU on the caller and move it to the desired devices on the callee if necessary.

Warning

TorchScript support in RPC is a prototype feature and subject to change. Since v1.5.0, torch.distributed.rpc supports calling TorchScript functions as RPC target functions, and this will help improve parallelism on the callee side as executing TorchScript functions does not require GIL.

torch.distributed.rpc.rpc_sync(to, func, args=None, kwargs=None, timeout=- 1.0)[source]¶

Make a blocking RPC call to run function func on worker to. RPC messages are sent and received in parallel to execution of Python code. This method is thread-safe.

Parameters

to (str or WorkerInfo or int) – name/rank/WorkerInfo of the destination worker.
func (callable) – a callable function, such as Python callables, builtin operators (e.g. add()) and annotated TorchScript functions.
args (tuple) – the argument tuple for the func invocation.
kwargs (dict) – is a dictionary of keyword arguments for the func invocation.
timeout (float, optional) – timeout in seconds to use for this RPC. If the RPC does not complete in this amount of time, an exception indicating it has timed out will be raised. A value of 0 indicates an infinite timeout, i.e. a timeout error will never be raised. If not provided, the default value set during initialization or with _set_rpc_timeout is used.

Returns

Returns the result of running func with args and kwargs.

Example::

Make sure that MASTER_ADDR and MASTER_PORT are set properly on both workers. Refer to init_process_group() API for more details. For example,

>>> export MASTER_ADDR=localhost
>>> export MASTER_PORT=5678

Then run the following code in two different processes:

>>> # On worker 0:
>>> import torch
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker0", rank=0, world_size=2)
>>> ret = rpc.rpc_sync("worker1", torch.add, args=(torch.ones(2), 3))
>>> rpc.shutdown()

>>> # On worker 1:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker1", rank=1, world_size=2)
>>> rpc.shutdown()

Below is an example of running a TorchScript function using RPC.

>>> # On both workers:
>>> @torch.jit.script
>>> def my_script_add(t1, t2):
>>>    return torch.add(t1, t2)

>>> # On worker 0:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker0", rank=0, world_size=2)
>>> ret = rpc.rpc_sync("worker1", my_script_add, args=(torch.ones(2), 3))
>>> rpc.shutdown()

>>> # On worker 1:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker1", rank=1, world_size=2)
>>> rpc.shutdown()

torch.distributed.rpc.rpc_async(to, func, args=None, kwargs=None, timeout=- 1.0)[source]¶

Make a non-blocking RPC call to run function func on worker to. RPC messages are sent and received in parallel to execution of Python code. This method is thread-safe. This method will immediately return a Future that can be awaited on.

Parameters

to (str or WorkerInfo or int) – name/rank/WorkerInfo of the destination worker.
func (callable) – a callable function, such as Python callables, builtin operators (e.g. add()) and annotated TorchScript functions.
args (tuple) – the argument tuple for the func invocation.
kwargs (dict) – is a dictionary of keyword arguments for the func invocation.
timeout (float, optional) – timeout in seconds to use for this RPC. If the RPC does not complete in this amount of time, an exception indicating it has timed out will be raised. A value of 0 indicates an infinite timeout, i.e. a timeout error will never be raised. If not provided, the default value set during initialization or with _set_rpc_timeout is used.

Returns

Returns a Future object that can be waited on. When completed, the return value of func on args and kwargs can be retrieved from the Future object.

Warning

Using GPU tensors as arguments or return values of func is not supported since we don’t support sending GPU tensors over the wire. You need to explicitly copy GPU tensors to CPU before using them as arguments or return values of func.

Warning

The rpc_async API does not copy storages of argument tensors until sending them over the wire, which could be done by a different thread depending on the RPC backend type. The caller should make sure that the contents of those tensors stay intact until the returned Future completes.

Example::

Make sure that MASTER_ADDR and MASTER_PORT are set properly on both workers. Refer to init_process_group() API for more details. For example,

>>> export MASTER_ADDR=localhost
>>> export MASTER_PORT=5678

Then run the following code in two different processes:

>>> # On worker 0:
>>> import torch
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker0", rank=0, world_size=2)
>>> fut1 = rpc.rpc_async("worker1", torch.add, args=(torch.ones(2), 3))
>>> fut2 = rpc.rpc_async("worker1", min, args=(1, 2))
>>> result = fut1.wait() + fut2.wait()
>>> rpc.shutdown()

>>> # On worker 1:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker1", rank=1, world_size=2)
>>> rpc.shutdown()

Below is an example of running a TorchScript function using RPC.

>>> # On both workers:
>>> @torch.jit.script
>>> def my_script_add(t1, t2):
>>>    return torch.add(t1, t2)

>>> # On worker 0:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker0", rank=0, world_size=2)
>>> fut = rpc.rpc_async("worker1", my_script_add, args=(torch.ones(2), 3))
>>> ret = fut.wait()
>>> rpc.shutdown()

>>> # On worker 1:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker1", rank=1, world_size=2)
>>> rpc.shutdown()

torch.distributed.rpc.remote(to, func, args=None, kwargs=None, timeout=- 1.0)[source]¶

Make a remote call to run func on worker to and return an RRef to the result value immediately. Worker to will be the owner of the returned RRef, and the worker calling remote is a user. The owner manages the global reference count of its RRef, and the owner RRef is only destructed when globally there are no living references to it.

Parameters

to (str or WorkerInfo or int) – name/rank/WorkerInfo of the destination worker.
func (callable) – a callable function, such as Python callables, builtin operators (e.g. add()) and annotated TorchScript functions.
args (tuple) – the argument tuple for the func invocation.
kwargs (dict) – is a dictionary of keyword arguments for the func invocation.
timeout (float, optional) – timeout in seconds for this remote call. If the creation of this RRef on worker to is not successfully processed on this worker within this timeout, then the next time there is an attempt to use the RRef (such as to_here()), a timeout will be raised indicating this failure. A value of 0 indicates an infinite timeout, i.e. a timeout error will never be raised. If not provided, the default value set during initialization or with _set_rpc_timeout is used.

Returns

A user RRef instance to the result value. Use the blocking API torch.distributed.rpc.RRef.to_here() to retrieve the result value locally.

Warning

The remote API does not copy storages of argument tensors until sending them over the wire, which could be done by a different thread depending on the RPC backend type. The caller should make sure that the contents of those tensors stay intact until the returned RRef is confirmed by the owner, which can be checked using the torch.distributed.rpc.RRef.confirmed_by_owner() API.

Warning

Errors such as timeouts for the remote API are handled on a best-effort basis. This means that when remote calls initiated by remote fail, such as with a timeout error, we take a best-effort approach to error handling. This means that errors are handled and set on the resulting RRef on an asynchronous basis. If the RRef has not been used by the application before this handling (such as to_here or fork call), then future uses of the RRef will appropriately raise errors. However, it is possible that the user application will use the RRef before the errors are handled. In this case, errors may not be raised as they have not yet been handled.

Example::

Make sure that MASTER_ADDR and MASTER_PORT are set properly on both workers. Refer to init_process_group() API for more details. For example,

>>> export MASTER_ADDR=localhost
>>> export MASTER_PORT=5678

Then run the following code in two different processes:

>>> # On worker 0:
>>> import torch
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker0", rank=0, world_size=2)
>>> rref1 = rpc.remote("worker1", torch.add, args=(torch.ones(2), 3))
>>> rref2 = rpc.remote("worker1", torch.add, args=(torch.ones(2), 1))
>>> x = rref1.to_here() + rref2.to_here()
>>> rpc.shutdown()

>>> # On worker 1:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker1", rank=1, world_size=2)
>>> rpc.shutdown()

Below is an example of running a TorchScript function using RPC.

>>> # On both workers:
>>> @torch.jit.script
>>> def my_script_add(t1, t2):
>>>    return torch.add(t1, t2)

>>> # On worker 0:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker0", rank=0, world_size=2)
>>> rref = rpc.remote("worker1", my_script_add, args=(torch.ones(2), 3))
>>> rref.to_here()
>>> rpc.shutdown()

>>> # On worker 1:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker1", rank=1, world_size=2)
>>> rpc.shutdown()

torch.distributed.rpc.get_worker_info(worker_name=None)[source]¶

Get WorkerInfo of a given worker name. Use this WorkerInfo to avoid passing an expensive string on every invocation.

Parameters: worker_name (str) – the string name of a worker. If None, return the the id of the current worker. (default None)
Returns: WorkerInfo instance for the given worker_name or WorkerInfo of the current worker if worker_name is None.

torch.distributed.rpc.shutdown(graceful=True)[source]¶

Perform a shutdown of the RPC agent, and then destroy the RPC agent. This stops the local agent from accepting outstanding requests, and shuts down the RPC framework by terminating all RPC threads. If graceful=True, this will block until all local and remote RPC processes reach this method and wait for all outstanding work to complete. Otherwise, if graceful=False, this is a local shutdown, and it does not wait for other RPC processes to reach this method.

Warning

For Future objects returned by rpc_async(), future.wait() should not be called after shutdown().

Parameters: graceful (bool) – Whether to do a graceful shutdown or not. If True, this will 1) wait until there is no pending system messages for UserRRefs and delete them; 2) block until all local and remote RPC processes have reached this method and wait for all outstanding work to complete.

Example::

Make sure that MASTER_ADDR and MASTER_PORT are set properly on both workers. Refer to init_process_group() API for more details. For example,

>>> export MASTER_ADDR=localhost
>>> export MASTER_PORT=5678

Then run the following code in two different processes:

>>> # On worker 0:
>>> import torch
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker0", rank=0, world_size=2)
>>> # do some work
>>> result = rpc.rpc_sync("worker1", torch.add, args=(torch.ones(1), 1))
>>> # ready to shutdown
>>> rpc.shutdown()

>>> # On worker 1:
>>> import torch.distributed.rpc as rpc
>>> rpc.init_rpc("worker1", rank=1, world_size=2)
>>> # wait for worker 0 to finish work, and then shutdown.
>>> rpc.shutdown()

class torch.distributed.rpc.WorkerInfo¶

A structure that encapsulates information of a worker in the system. Contains the name and ID of the worker. This class is not meant to be constructed directly, rather, an instance can be retrieved through get_worker_info() and the result can be passed in to functions such as rpc_sync(), rpc_async(), remote() to avoid copying a string on every invocation.

property id¶: Globally unique id to identify the worker.

property name¶: The name of the worker.

The RPC package also provides decorators which allow applications to specify how a given function should be treated on the callee side.

torch.distributed.rpc.functions.async_execution(fn)[source]¶

A decorator for a function indicating that the return value of the function is guaranteed to be a Future object and this function can run asynchronously on the RPC callee. More specifically, the callee extracts the Future returned by the wrapped function and installs subsequent processing steps as a callback to that Future. The installed callback will read the value from the Future when completed and send the value back as the RPC response. That also means the returned Future only exists on the callee side and is never sent through RPC. This decorator is useful when the wrapped function’s (fn) execution needs to pause and resume due to, e.g., containing rpc_async() or waiting for other signals.

Note

To enable asynchronous execution, applications must pass the function object returned by this decorator to RPC APIs. If RPC detected attributes installed by this decorator, it knows that this function returns a Future object and will handle that accordingly. However, this does not mean this decorator has to be outmost one when defining a function. For example, when combined with @staticmethod or @classmethod, @rpc.functions.async_execution needs to be the inner decorator to allow the target function be recognized as a static or class function. This target function can still execute asynchronously because, when accessed, the static or class method preserves attributes installed by @rpc.functions.async_execution.

Example::

The returned Future object can come from rpc_async(), then(), or Future constructor. The example below shows directly using the Future returned by then().

>>> from torch.distributed import rpc
>>>
>>> # omitting setup and shutdown RPC
>>>
>>> # On all workers
>>> @rpc.functions.async_execution
>>> def async_add_chained(to, x, y, z):
>>>     # This function runs on "worker1" and returns immediately when
>>>     # the callback is installed through the `then(cb)` API. In the
>>>     # mean time, the `rpc_async` to "worker2" can run concurrently.
>>>     # When the return value of that `rpc_async` arrives at
>>>     # "worker1", "worker1" will run the lambda function accordingly
>>>     # and set the value for the previously returned `Future`, which
>>>     # will then trigger RPC to send the result back to "worker0".
>>>     return rpc.rpc_async(to, torch.add, args=(x, y)).then(
>>>         lambda fut: fut.wait() + z
>>>     )
>>>
>>> # On worker0
>>> ret = rpc.rpc_sync(
>>>     "worker1",
>>>     async_add_chained,
>>>     args=("worker2", torch.ones(2), 1, 1)
>>> )
>>> print(ret)  # prints tensor([3., 3.])

When combined with TorchScript decorators, this decorator must be the outmost one.

>>> from torch import Tensor
>>> from torch.futures import Future
>>> from torch.distributed import rpc
>>>
>>> # omitting setup and shutdown RPC
>>>
>>> # On all workers
>>> @torch.jit.script
>>> def script_add(x: Tensor, y: Tensor) -> Tensor:
>>>     return x + y
>>>
>>> @rpc.functions.async_execution
>>> @torch.jit.script
>>> def async_add(to: str, x: Tensor, y: Tensor) -> Future[Tensor]:
>>>     return rpc.rpc_async(to, script_add, (x, y))
>>>
>>> # On worker0
>>> ret = rpc.rpc_sync(
>>>     "worker1",
>>>     async_add,
>>>     args=("worker2", torch.ones(2), 1)
>>> )
>>> print(ret)  # prints tensor([2., 2.])

When combined with static or class method, this decorator must be the inner one.

>>> from torch.distributed import rpc
>>>
>>> # omitting setup and shutdown RPC
>>>
>>> # On all workers
>>> class AsyncExecutionClass:
>>>
>>>     @staticmethod
>>>     @rpc.functions.async_execution
>>>     def static_async_add(to, x, y, z):
>>>         return rpc.rpc_async(to, torch.add, args=(x, y)).then(
>>>             lambda fut: fut.wait() + z
>>>         )
>>>
>>>     @classmethod
>>>     @rpc.functions.async_execution
>>>     def class_async_add(cls, to, x, y, z):
>>>         ret_fut = torch.futures.Future()
>>>         rpc.rpc_async(to, torch.add, args=(x, y)).then(
>>>             lambda fut: ret_fut.set_result(fut.wait() + z)
>>>         )
>>>         return ret_fut
>>>
>>>     @rpc.functions.async_execution
>>>     def bound_async_add(self, to, x, y, z):
>>>         return rpc.rpc_async(to, torch.add, args=(x, y)).then(
>>>             lambda fut: fut.wait() + z
>>>         )
>>>
>>> # On worker0
>>> ret = rpc.rpc_sync(
>>>     "worker1",
>>>     AsyncExecutionClass.static_async_add,
>>>     args=("worker2", torch.ones(2), 1, 2)
>>> )
>>> print(ret)  # prints tensor([4., 4.])
>>>
>>> ret = rpc.rpc_sync(
>>>     "worker1",
>>>     AsyncExecutionClass.class_async_add,
>>>     args=("worker2", torch.ones(2), 1, 2)
>>> )
>>> print(ret)  # prints tensor([4., 4.])

This decorator also works with RRef helpers, i.e., . torch.distributed.rpc.RRef.rpc_sync(), torch.distributed.rpc.RRef.rpc_async(), and torch.distributed.rpc.RRef.remote().

>>> from torch.distributed import rpc
>>>
>>> # reuse the AsyncExecutionClass class above
>>> rref = rpc.remote("worker1", AsyncExecutionClass)
>>> ret = rref.rpc_sync().static_async_add("worker2", torch.ones(2), 1, 2)
>>> print(ret)  # prints tensor([4., 4.])
>>>
>>> rref = rpc.remote("worker1", AsyncExecutionClass)
>>> ret = rref.rpc_async().static_async_add("worker2", torch.ones(2), 1, 2).wait()
>>> print(ret)  # prints tensor([4., 4.])
>>>
>>> rref = rpc.remote("worker1", AsyncExecutionClass)
>>> ret = rref.remote().static_async_add("worker2", torch.ones(2), 1, 2).to_here()
>>> print(ret)  # prints tensor([4., 4.])

Backends¶

The RPC module can leverage different backends to perform the communication between the nodes. The backend to be used can be specified in the init_rpc() function, by passing a certain value of the BackendType enum. Regardless of what backend is used, the rest of the RPC API won’t change. Each backend also defines its own subclass of the RpcBackendOptions class, an instance of which can also be passed to init_rpc() to configure the backend’s behavior.

class torch.distributed.rpc.BackendType(value)¶

An enum class of available backends.

PyTorch ships with a builtin BackendType.TENSORPIPE backend. Additional ones can be registered using the register_backend() function.

class torch.distributed.rpc.RpcBackendOptions¶

An abstract structure encapsulating the options passed into the RPC backend. An instance of this class can be passed in to init_rpc() in order to initialize RPC with specific configurations, such as the RPC timeout and init_method to be used.

property init_method¶: URL specifying how to initialize the process group. Default is env://

property rpc_timeout¶: A float indicating the timeout to use for all RPCs. If an RPC does not complete in this timeframe, it will complete with an exception indicating that it has timed out.

TensorPipe Backend¶

The TensorPipe agent, which is the default, leverages the TensorPipe library, which provides a natively point-to-point communication primitive specifically suited for machine learning that fundamentally addresses some of the limitations of Gloo. Compared to Gloo, it has the advantage of being asynchronous, which allows a large number of transfers to occur simultaneously, each at their own speed, without blocking each other. It will only open pipes between pairs of nodes when needed, on demand, and when one node fails only its incident pipes will be closed, while all other ones will keep working as normal. In addition, it is able to support multiple different transports (TCP, of course, but also shared memory, NVLink, InfiniBand, …) and can automatically detect their availability and negotiate the best transport to use for each pipe.

The TensorPipe backend has been introduced in PyTorch v1.6 and is being actively developed. At the moment, it only supports CPU tensors, with GPU support coming soon. It comes with a TCP-based transport, just like Gloo. It is also able to automatically chunk and multiplex large tensors over multiple sockets and threads in order to achieve very high bandwidths. The agent will be able to pick the best transport on its own, with no intervention required.

Example:

>>> import os
>>> from torch.distributed import rpc
>>> os.environ['MASTER_ADDR'] = 'localhost'
>>> os.environ['MASTER_PORT'] = '29500'
>>>
>>> rpc.init_rpc(
>>>     "worker1",
>>>     rank=0,
>>>     world_size=2,
>>>     rpc_backend_options=rpc.TensorPipeRpcBackendOptions(
>>>         num_worker_threads=8,
>>>         rpc_timeout=20 # 20 second timeout
>>>     )
>>> )
>>>
>>> # omitting init_rpc invocation on worker2

class torch.distributed.rpc.TensorPipeRpcBackendOptions(*, num_worker_threads=16, rpc_timeout=60.0, init_method='env://', device_maps=None, devices=None, _transports=None, _channels=None)[source]¶

The backend options for TensorPipeAgent, derived from RpcBackendOptions.

Parameters

num_worker_threads (int, optional) – The number of threads in the thread-pool used by TensorPipeAgent to execute requests (default: 16).
rpc_timeout (float, optional) – The default timeout, in seconds, for RPC requests (default: 60 seconds). If the RPC has not completed in this timeframe, an exception indicating so will be raised. Callers can override this timeout for individual RPCs in rpc_sync() and rpc_async() if necessary.
init_method (str, optional) – The URL to initialize the distributed store used for rendezvous. It takes any value accepted for the same argument of init_process_group() (default: env://).
device_maps (Dict[str, Dict], optional) – Device placement mappings from this worker to the callee. Key is the callee worker name and value the dictionary (Dict of int, str, or torch.device) that maps this worker’s devices to the callee worker’s devices. (default: None)
devices (List[int, str, or torch.device], optional) – all local CUDA devices used by RPC agent. By Default, it will be initialized to all local devices from its own device_maps and corresponding devices from its peers’ device_maps. When processing CUDA RPC requests, the agent will properly synchronize CUDA streams for all devices in this List.

property device_maps¶: The device map locations.

property devices¶: All devices used by the local agent.

property init_method¶: URL specifying how to initialize the process group. Default is env://

property num_worker_threads¶: The number of threads in the thread-pool used by TensorPipeAgent to execute requests.

property rpc_timeout¶: A float indicating the timeout to use for all RPCs. If an RPC does not complete in this timeframe, it will complete with an exception indicating that it has timed out.

set_device_map(to, device_map)[source]¶

Set device mapping between each RPC caller and callee pair. This function can be called multiple times to incrementally add device placement configurations.

Parameters

worker_name (str) – Callee name.
device_map (Dict of python:int, str, or torch.device) – Device placement mappings from this worker to the callee. This map must be invertible.

Example::

>>> # both workers
>>> def add(x, y):
>>>     print(x)  # tensor([1., 1.], device='cuda:1')
>>>     return x + y, (x + y).to(2)
>>>
>>> # on worker 0
>>> options = TensorPipeRpcBackendOptions(
>>>     num_worker_threads=8,
>>>     device_maps={"worker1": {0: 1}}
>>>     # maps worker0's cuda:0 to worker1's cuda:1
>>> )
>>> options.set_device_map("worker1", {1: 2})
>>> # maps worker0's cuda:1 to worker1's cuda:2
>>>
>>> rpc.init_rpc(
>>>     "worker0",
>>>     rank=0,
>>>     world_size=2,
>>>     backend=rpc.BackendType.TENSORPIPE,
>>>     rpc_backend_options=options
>>> )
>>>
>>> x = torch.ones(2)
>>> rets = rpc.rpc_sync("worker1", add, args=(x.to(0), 1))
>>> # The first argument will be moved to cuda:1 on worker1. When
>>> # sending the return value back, it will follow the invert of
>>> # the device map, and hence will be moved back to cuda:0 and
>>> # cuda:1 on worker0
>>> print(rets[0])  # tensor([2., 2.], device='cuda:0')
>>> print(rets[1])  # tensor([2., 2.], device='cuda:1')

set_devices(devices)[source]¶

Set local devices used by the TensorPipe RPC agent. When processing CUDA RPC requests, the TensorPipe RPC agent will properly synchronize CUDA streams for all devices in this List.

Parameters: devices (List of python:int, str, or torch.device) – local devices used by the TensorPipe RPC agent.

RRef¶

Warning

RRefs are not currently supported when using CUDA tensors

An RRef (Remote REFerence) is a reference to a value of some type T (e.g. Tensor) on a remote worker. This handle keeps the referenced remote value alive on the owner, but there is no implication that the value will be transferred to the local worker in the future. RRefs can be used in multi-machine training by holding references to nn.Modules that exist on other workers, and calling the appropriate functions to retrieve or modify their parameters during training. See Remote Reference Protocol for more details.

class torch.distributed.rpc.RRef[source]¶

backward(self: torch._C._distributed_rpc.PyRRef, dist_autograd_ctx_id: int = - 1, retain_graph: bool = False) → None ¶

Runs the backward pass using the RRef as the root of the backward pass. If dist_autograd_ctx_id is provided, we perform a distributed backward pass using the provided ctx_id starting from the owner of the RRef. In this case, get_gradients() should be used to retrieve the gradients. If dist_autograd_ctx_id is None, it is assumed that this is a local autograd graph and we only perform a local backward pass. In the local case, the node calling this API has to be the owner of the RRef. The value of the RRef is expected to be a scalar Tensor.

Parameters

dist_autograd_ctx_id (int, optional) – The distributed autograd context id for which we should retrieve the gradients (default: -1).
retain_graph (bool, optional) – If False, the graph used to compute the grad will be freed. Note that in nearly all cases setting this option to True is not needed and often can be worked around in a much more efficient way. Usually, you need to set this to True to run backward multiple times (default: False).

Example::

>>> import torch.distributed.autograd as dist_autograd
>>> with dist_autograd.context() as context_id:
>>>     rref.backward(context_id)

confirmed_by_owner(self: torch._C._distributed_rpc.PyRRef) → bool ¶: Returns whether this RRef has been confirmed by the owner. OwnerRRef always returns true, while UserRRef only returns true when the owner knowns about this UserRRef.

is_owner(self: torch._C._distributed_rpc.PyRRef) → bool ¶: Returns whether or not the current node is the owner of this RRef.

local_value(self: torch._C._distributed_rpc.PyRRef) → object ¶: If the current node is the owner, returns a reference to the local value. Otherwise, throws an exception.

owner(self: torch._C._distributed_rpc.PyRRef) → torch._C._distributed_rpc.WorkerInfo¶: Returns worker information of the node that owns this RRef.

owner_name(self: torch._C._distributed_rpc.PyRRef) → str ¶: Returns worker name of the node that owns this RRef.

remote(self: torch._C._distributed_rpc.PyRRef, timeout: float = - 1.0) → object ¶

Create a helper proxy to easily launch a remote using the owner of the RRef as the destination to run functions on the object referenced by this RRef. More specifically, rref.remote().func_name(*args, **kwargs) is the same as the following:

>>> def run(rref, func_name, args, kwargs):
>>>   return getattr(rref.local_value(), func_name)(*args, **kwargs)
>>>
>>> rpc.remote(rref.owner(), run, args=(rref, func_name, args, kwargs))

Parameters: timeout (float, optional) – Timeout for rref.remote(). If the creation of this RRef is not successfully completed within the timeout, then the next time there is an attempt to use the RRef (such as to_here), a timeout will be raised. If not provided, the default RPC timeout will be used. Please see rpc.remote() for specific timeout semantics for RRef.

Example::

>>> from torch.distributed import rpc
>>> rref = rpc.remote("worker1", torch.add, args=(torch.zeros(2, 2), 1))
>>> rref.remote().size().to_here()  # returns torch.Size([2, 2])
>>> rref.remote().view(1, 4).to_here()  # returns tensor([[1., 1., 1., 1.]])

rpc_async(self: torch._C._distributed_rpc.PyRRef, timeout: float = - 1.0) → object ¶

Create a helper proxy to easily launch an rpc_async using the owner of the RRef as the destination to run functions on the object referenced by this RRef. More specifically, rref.rpc_async().func_name(*args, **kwargs) is the same as the following:

>>> def run(rref, func_name, args, kwargs):
>>>   return getattr(rref.local_value(), func_name)(*args, **kwargs)
>>>
>>> rpc.rpc_async(rref.owner(), run, args=(rref, func_name, args, kwargs))

Parameters: timeout (float, optional) – Timeout for rref.rpc_async(). If the call does not complete within this timeframe, an exception indicating so will be raised. If this argument is not provided, the default RPC timeout will be used.

Example::

>>> from torch.distributed import rpc
>>> rref = rpc.remote("worker1", torch.add, args=(torch.zeros(2, 2), 1))
>>> rref.rpc_async().size().wait()  # returns torch.Size([2, 2])
>>> rref.rpc_async().view(1, 4).wait()  # returns tensor([[1., 1., 1., 1.]])

rpc_sync(self: torch._C._distributed_rpc.PyRRef, timeout: float = - 1.0) → object ¶

Create a helper proxy to easily launch an rpc_sync using the owner of the RRef as the destination to run functions on the object referenced by this RRef. More specifically, rref.rpc_sync().func_name(*args, **kwargs) is the same as the following:

>>> def run(rref, func_name, args, kwargs):
>>>   return getattr(rref.local_value(), func_name)(*args, **kwargs)
>>>
>>> rpc.rpc_sync(rref.owner(), run, args=(rref, func_name, args, kwargs))

Parameters: timeout (float, optional) – Timeout for rref.rpc_sync(). If the call does not complete within this timeframe, an exception indicating so will be raised. If this argument is not provided, the default RPC timeout will be used.

Example::

>>> from torch.distributed import rpc
>>> rref = rpc.remote("worker1", torch.add, args=(torch.zeros(2, 2), 1))
>>> rref.rpc_sync().size()  # returns torch.Size([2, 2])
>>> rref.rpc_sync().view(1, 4)  # returns tensor([[1., 1., 1., 1.]])

to_here(self: torch._C._distributed_rpc.PyRRef, timeout: float = - 1.0) → object ¶

Blocking call that copies the value of the RRef from the owner to the local node and returns it. If the current node is the owner, returns a reference to the local value.

Parameters: timeout (float, optional) – Timeout for to_here. If the call does not complete within this timeframe, an exception indicating so will be raised. If this argument is not provided, the default RPC timeout (60s) will be used.

More Information about RRef

Remote Reference Protocol

RemoteModule¶

Warning

RemoteModule is not currently supported when using CUDA tensors

RemoteModule is an easy way to create an nn.Module remotely on a different process. The actual module resides on a remote host, but the local host has a handle to this module and invoke this module similar to a regular nn.Module. The invocation however incurs RPC calls to the remote end and can be performed asynchronously if needed via additional APIs supported by RemoteModule.

class torch.distributed.nn.api.remote_module.RemoteModule(remote_device, module_cls, args=None, kwargs=None)[source]¶

A RemoteModule instance can only be created after RPC initialization. It creates a user-specified module on a specified remote node. It behaves like a regular nn.Module except that the forward method is executed on the remote node. It takes care of autograd recording to ensure the backward pass propogates gradients back to the corresponding remote module.

It generates two methods forward_async and forward based on the signature of the forward method of module_cls. forward_async runs asynchronously and returns a Future. The arguments of forward_async and forward are the same as the forward method of the module returned by the module_cls.

For example, if module_cls returns an instance of nn.Linear, that has forward method signature: def forward(input: Tensor) -> Tensor:, the generated RemoteModule will have 2 methods with the signatures:

def forward(input: Tensor) -> Tensor:

def forward_async(input: Tensor) -> Future[Tensor]:

Parameters

remote_device (str) – Device on the destination worker where we’d like to place this module. The format should be “<workername>/<device>”, where the device field can be parsed as torch.device type. E.g., “trainer0/cpu”, “trainer0”, “ps0/cuda:0”. In addition, the device field can be optional and the default value is “cpu”.

module_cls (nn.Module) –

Class for the module to be created remotely. For example,

>>> class MyModule(nn.Module):
>>>     def forward(input):
>>>         return input + 1
>>>
>>> module_cls = MyModule

args (Sequence, optional) – args to be passed to module_cls.
kwargs (Dict, optional) – kwargs to be passed to module_cls.

Returns

A remote module instance which wraps the Module created by the user-provided module_cls, it has a blocking forward method and an asynchronous forward_async method that returns a future of the forward call on the user-provided module on the remote side.

Example::

Run the following code in two different processes:

>>> # On worker 0:
>>> import torch
>>> import torch.distributed.rpc as rpc
>>> from torch import nn, Tensor
>>> from torch.distributed.nn.api.remote_module import RemoteModule
>>>
>>> rpc.init_rpc("worker0", rank=0, world_size=2)
>>> remote_linear_module = RemoteModule(
>>>     "worker1/cpu", nn.Linear, args=(20, 30),
>>> )
>>> input = torch.randn(128, 20)
>>> ret_fut = remote_linear_module.forward_async(input)
>>> ret = ret_fut.wait()
>>> rpc.shutdown()

>>> # On worker 1:
>>> import torch
>>> import torch.distributed.rpc as rpc
>>>
>>> rpc.init_rpc("worker1", rank=1, world_size=2)
>>> rpc.shutdown()

Furthermore, a more practical example that is combined with DistributedDataParallel (DDP) can be found in this tutorial.

get_module_rref()¶: Returns an RRef (RRef[nn.Module]) pointing to the remote module.

remote_parameters(recurse=True)¶

Returns a list of RRef pointing to the remote module’s parameters. This can typically be used in conjuction with DistributedOptimizer.

Parameters: recurse (bool) – if True, then returns parameters of the remote module and all submodules of the remote module. Otherwise, returns only parameters that are direct members of the remote module.
Returns: A list of RRef (List[RRef[nn.Parameter]]) to remote module’s parameters.

Distributed Autograd Framework¶

Warning

Distributed autograd is not currently supported when using CUDA tensors

This module provides an RPC-based distributed autograd framework that can be used for applications such as model parallel training. In short, applications may send and receive gradient recording tensors over RPC. In the forward pass, we record when gradient recording tensors are sent over RPC and during the backward pass we use this information to perform a distributed backward pass using RPC. For more details see Distributed Autograd Design.

torch.distributed.autograd.backward(context_id: int, roots: List[Tensor], retain_graph=False) → None ¶

Kicks off the distributed backward pass using the provided roots. This currently implements the FAST mode algorithm which assumes all RPC messages sent in the same distributed autograd context across workers would be part of the autograd graph during the backward pass.

We use the provided roots to discover the autograd graph and compute appropriate dependencies. This method blocks until the entire autograd computation is done.

We accumulate the gradients in the appropriate torch.distributed.autograd.context on each of the nodes. The autograd context to be used is looked up given the context_id that is passed in when torch.distributed.autograd.backward() is called. If there is no valid autograd context corresponding to the given ID, we throw an error. You can retrieve the accumulated gradients using the get_gradients() API.

Parameters

context_id (int) – The autograd context id for which we should retrieve the gradients.
roots (list) – Tensors which represent the roots of the autograd computation. All the tensors should be scalars.
retain_graph (bool, optional) – If False, the graph used to compute the grad will be freed. Note that in nearly all cases setting this option to True is not needed and often can be worked around in a much more efficient way. Usually, you need to set this to True to run backward multiple times.

Example::

>>> import torch.distributed.autograd as dist_autograd
>>> with dist_autograd.context() as context_id:
>>>     pred = model.forward()
>>>     loss = loss_func(pred, loss)
>>>     dist_autograd.backward(context_id, loss)

class torch.distributed.autograd.context[source]¶

Context object to wrap forward and backward passes when using distributed autograd. The context_id generated in the with statement is required to uniquely identify a distributed backward pass on all workers. Each worker stores metadata associated with this context_id, which is required to correctly execute a distributed autograd pass.

Example::

>>> import torch.distributed.autograd as dist_autograd
>>> with dist_autograd.context() as context_id:
>>>   t1 = torch.rand((3, 3), requires_grad=True)
>>>   t2 = torch.rand((3, 3), requires_grad=True)
>>>   loss = rpc.rpc_sync("worker1", torch.add, args=(t1, t2)).sum()
>>>   dist_autograd.backward(context_id, [loss])

torch.distributed.autograd.get_gradients(context_id: int) → Dict[Tensor, Tensor]¶

Retrieves a map from Tensor to the appropriate gradient for that Tensor accumulated in the provided context corresponding to the given context_id as part of the distributed autograd backward pass.

Parameters: context_id (int) – The autograd context id for which we should retrieve the gradients.
Returns: A map where the key is the Tensor and the value is the associated gradient for that Tensor.

Example::

>>> import torch.distributed.autograd as dist_autograd
>>> with dist_autograd.context() as context_id:
>>>     t1 = torch.rand((3, 3), requires_grad=True)
>>>     t2 = torch.rand((3, 3), requires_grad=True)
>>>     loss = t1 + t2
>>>     dist_autograd.backward(context_id, [loss.sum()])
>>>     grads = dist_autograd.get_gradients(context_id)
>>>     print(grads[t1])
>>>     print(grads[t2])

More Information about RPC Autograd

Distributed Autograd Design

Distributed Optimizer¶

Warning

Distributed optimizer is not currently supported when using CUDA tensors

torch.distributed.optim exposes DistributedOptimizer, which takes a list of remote parameters (RRef) and runs the optimizer locally on the workers where the parameters live. The distributed optimizer can use any of the local optimizer Base class to apply the gradients on each worker.

class torch.distributed.optim.DistributedOptimizer(optimizer_class, params_rref, *args, **kwargs)[source]¶

DistributedOptimizer takes remote references to parameters scattered across workers and applies the given optimizer locally for each parameter.

This class uses get_gradients() in order to retrieve the gradients for specific parameters.

Concurrent calls to step(), either from the same or different clients, will be serialized on each worker – as each worker’s optimizer can only work on one set of gradients at a time. However, there is no guarantee that the full forward-backward-optimizer sequence will execute for one client at a time. This means that the gradients being applied may not correspond to the latest forward pass executed on a given worker. Also, there is no guaranteed ordering across workers.

DistributedOptimizer creates the local optimizer with TorchScript enabled by default, so that optimizer updates are not blocked by the Python Global Interpreter Lock (GIL) in the case of multithreaded training (e.g. Distributed Model Parallel). This feature is currently enabled for most optimizers. You can also follow the recipe in PyTorch tutorials to enable TorchScript support for your own custom optimizers.

Parameters

optimizer_class (optim.Optimizer) – the class of optimizer to instantiate on each worker.
params_rref (list[RRef]) – list of RRefs to local or remote parameters to optimize.
args – arguments to pass to the optimizer constructor on each worker.
kwargs – arguments to pass to the optimizer constructor on each worker.

Example::

>>> import torch.distributed.autograd as dist_autograd
>>> import torch.distributed.rpc as rpc
>>> from torch import optim
>>> from torch.distributed.optim import DistributedOptimizer
>>>
>>> with dist_autograd.context() as context_id:
>>>   # Forward pass.
>>>   rref1 = rpc.remote("worker1", torch.add, args=(torch.ones(2), 3))
>>>   rref2 = rpc.remote("worker1", torch.add, args=(torch.ones(2), 1))
>>>   loss = rref1.to_here() + rref2.to_here()
>>>
>>>   # Backward pass.
>>>   dist_autograd.backward(context_id, [loss.sum()])
>>>
>>>   # Optimizer.
>>>   dist_optim = DistributedOptimizer(
>>>      optim.SGD,
>>>      [rref1, rref2],
>>>      lr=0.05,
>>>   )
>>>   dist_optim.step(context_id)

step(context_id)[source]¶

Performs a single optimization step.

This will call torch.optim.Optimizer.step() on each worker containing parameters to be optimized, and will block until all workers return. The provided context_id will be used to retrieve the corresponding context that contains the gradients that should be applied to the parameters.

Parameters: context_id – the autograd context id for which we should run the optimizer step.

Design Notes¶

The distributed autograd design note covers the design of the RPC-based distributed autograd framework that is useful for applications such as model parallel training.

Distributed Autograd Design

The RRef design note covers the design of the RRef (Remote REFerence) protocol used to refer to values on remote workers by the framework.

Remote Reference Protocol

Tutorials¶

The RPC tutorials introduce users to the RPC framework, provide several example applications using torch.distributed.rpc APIs, and demonstrate how to use the profiler to profile RPC-based workloads.

Distributed RPC Framework¶

Basics¶

RPC¶

Backends¶

TensorPipe Backend¶

RRef¶

RemoteModule¶

Distributed Autograd Framework¶

Distributed Optimizer¶

Design Notes¶

Tutorials¶

Docs

Tutorials

Resources